The configuration is stored separately to the Vm in the VMM. The failure
to store the config was preventing the VM from shutting down correctly
as Vmm::vm_delete() checks for the presence of the config.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The live migration support added use of this ioctl but it wasn't
included in the permitted list.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This interface is used by the vCPU thread to delegate responsibility for
handling MMIO/PIO operations and to support different approaches than a
VM exit.
During profiling I found that we were spending 13.75% of the boot CPU
uage acquiring access to the object holding the VmmOps via
ArcSwap::load_full()
13.75% 6.02% vcpu0 cloud-hypervisor [.] arc_swap::ArcSwapAny<T,S>::load_full
|
---arc_swap::ArcSwapAny<T,S>::load_full
|
--13.43%--<hypervisor::kvm::KvmVcpu as hypervisor::cpu::Vcpu>::run
std::sys_common::backtrace::__rust_begin_short_backtrace
core::ops::function::FnOnce::call_once{{vtable-shim}}
std::sys::unix:🧵:Thread:🆕:thread_start
However since the object implementing VmmOps does not need to be mutable
and it is only used from the vCPU side we can change the ownership to
being a simple Arc<> that is passed in when calling create_vcpu().
This completely removes the above CPU usage from subsequent profiles.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Add config parameter to --disk called "_disable_io_uring" (the
underscore prefix indicating it is not for public consumpion.) Use this
option to disable io_uring if it would otherwise be used.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Now the VM is paused/resumed by the migration process itself.
0. The guest configuration is sent to the destination
1. Dirty page log tracking is started by start_memory_dirty_log()
2. All guest memory is sent to the destination
3. Up to 5 attempts are made to send the dirty guest memory to the
destination...
4. ...before the VM is paused
5. One last set of dirty pages is sent to the destination
6. The guest is snapshotted and sent to the destination
7. When the migration is completed the destination unpauses the received
VM.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This allows code running in the VMM to access the VM's MemoryManager's
functionality for managing the dirty log including resetting it but also
generating a table.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Prior to sending the memory the full state is not needed only the
configuration. This is sufficient to create the appropriate structures
in the guest and have the memory allocations ready for filling.
Update the protocol documentation to add a separate config step and move
the state to after the memory is transferred. As the VM is created in a
separate step to restoring it the requires a slightly different
constructor as well as saving the VM object for the subsequent commands.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
In order to do this we must extend the MemoryManager API to add the
ability to specify the tracking of the dirty pages when creating the
userspace mappings and also keep track of the userspace mappings that
have been created for RAM regions.
Currently the dirty pages are collected into ranges based on a block
level of 64 pages. The algorithm could be tweaked to create smaller
ranges but for now if any page in the block of 64 is dirty the whole
block is added to the range.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
While the addressable space size reduction of 4k in necessary due to
the Linux bug, the 64k alignment of the addressable space size is
required by Windows. This patch satisfies both.
Signed-off-by: Anatol Belski <anbelski@linux.microsoft.com>
This is tested by:
Source VMM:
target/debug/cloud-hypervisor --kernel ~/src/linux/vmlinux \
--pmem file=~/workloads/focal.raw --cpus boot=1 \
--memory size=2048M \
--cmdline"root=/dev/pmem0p1 console=ttyS0" --serial tty --console off \
--api-socket=/tmp/api1 -v
Destination VMM:
target/debug/cloud-hypervisor --api-socket=/tmp/api2 -v
And the following commands:
target/debug/ch-remote --api-socket=/tmp/api1 pause
target/debug/ch-remote --api-socket=/tmp/api2 receive-migration unix:/tmp/foo &
target/debug/ch-remote --api-socket=/tmp/api1 send-migration unix:/tmp/foo
target/debug/ch-remote --api-socket=/tmp/api2 resume
The VM is then responsive on the destination VMM.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This allows the code to be reused when creating the VM from a snapshot
when doing VM migration.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Add API entry points with stub implementation for sending and receiving
a VM from one VMM to another.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Due to a known limitation in OpenAPITools/openapi-generator tool,
it's impossible to send go zero types, like false and 0 to
cloud-hypervisor because `omitempty` is added if a field is not
required.
Set cache_size, dax, num_queues and queue_size as required to remove
`omitempty` from the json tag.
fixes#1961
Signed-off-by: Julio Montes <julio.montes@intel.com>
This also removes the need to lookup up the "exe" symlink for finding
the VMM executable path.
Fixes: #1925
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The logic to handle AArch64 system event was: SHUTDOWN and RESET were
all treated as RESET.
Now we handle them differently:
- RESET event will trigger Vmm::vm_reboot(),
- SHUTDOWN event will trigger Vmm::vm_shutdown().
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
Now Vcpu::run() returns a boolean value to VcpuManager, indicating
whether the VM is going to reboot (false) or just continue (true).
Moving the handling of hypervisor VCPU run result from Vcpu to
VcpuManager gives us the flexibility to handle more scenarios like
shutting down on AArch64.
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
Rather than filling the guest memory from a file at the point of the the
guest memory region being created instead fill from the file later. This
simplifies the region creation code but also adds flexibility for
sourcing the guest memory from a source other than an on disk file.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
As a mirror of bdbea19e23 which ensured
that GuestMemoryMmap::read_exact_from() was used to read all the file to
the region ensure that all the guest memory region is written to disk.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This gives a nicer user experience and this error can now be used as the
source for other errors based off this.
See: #1910
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Any occurrence of of a variable containing `ext_region` is replaced with
the less confusing name `saved_region`. The point is to clearly identify
the memory regions that might have been saved during a snapshot, while
the `ext` standing for `external` was pretty unclear.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
In the context of saving the memory regions content through snapshot,
using the term "backing file" brings confusion with the actual backing
file that might back the memory mapping.
To avoid such conflicting naming, the 'backing_file' field from the
MemoryRegion structure gets replaced with 'content', as this is
designating the potential file containing the memory region data.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Use GuestRegionMmap::read_exact_from() to ensure that all of the file is
read into the guest. This addresses an issue where
GuestRegionMmap::read_from() was only copying the first 2GiB of the
memory and so lead to snapshot-restore was failing when the guest RAM
was 2GiB or greater.
This change also propagates any error from the copying upwards.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
When restoring if a region of RAM is backed by anonymous memory i.e from
memfd_create() then copy the contents of the ram from the file that has
been saved to disk.
Previously the code would map the memory from that file into the guest
using a MAP_PRIVATE mapping. This has the effect of
minimising the restore time but provides an issue where the restored VM
does not have the same structure as the snapshotted VM, in particular
memory is backed by files in the restored VM that were anonymously
backed in the original.
This creates two problems:
* The snapshot data is mapped from files for the pages of the guest
which prevents the storage from being reclaimed.
* When snapshotting again the guest memory will not be correctly saved
as it will have looked like it was backed by a file so it will not be
written to disk but as it is a MAP_PRIVATE mapping the changes will
never be written to the disk again. This results in incorrect
behaviour.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The standalone `--balloon` parameter being fully functional at this
point, we can get rid of the balloon options from the --memory
parameter.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Now that we have a new dedicated way of asking for a balloon through the
CLI and the REST API, we can move all the balloon code to the device
manager. This allows us to simplify the memory manager, which is already
quite complex.
It also simplifies the behavior of the balloon resizing command. Instead
of providing the expected size for the RAM, which is complex when memory
zones are involved, it now expects the balloon size. This is a much more
straightforward behavior as it really resizes the balloon to the desired
size. Additionally to the simplication, the benefit of this approach is
that it does not need to be tied to the memory manager at all.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This introduces a new way of defining the virtio-balloon device. Instead
of going through the --memory parameter, the idea is to consider balloon
as a standalone virtio device.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The snasphot/restore feature is not working because some CPU states are
not properly saved, which means they can't be restored later on.
First thing, we ensure the CPUID is stored so that it can be properly
restored later. The code is simplified and pushed down to the hypervisor
crate.
Second thing, we identify for each vCPU if the Hyper-V SynIC device is
emulated or not. In case it is, that means some specific MSRs will be
set by the guest. These MSRs must be saved in order to properly restore
the VM.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The watchdog device is created through the "--watchdog" parameter. At
most a single watchdog can be created per VM.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
Before Virtio-mmio was removed, we passed an optional PCI space address
parameter to AArch64 code for generating FDT. The address is none if the
transport is MMIO.
Now Virtio-PCI is the only option, the parameter is mandatory.
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
Virtio-mmio is removed, now virtio-pci is the only option for virtio
transport layer. We use MSI for PCI device interrupt. While GICv2, the
legacy interrupt controller, doesn't support MSI. So GICv2 is not very
practical for Cloud-hypervisor, we can remove it.
Signed-off-by: Michael Zhao <michael.zhao@arm.com>
When shutting down a VM using VFIO, the following error has been
detected:
vfio-ioctls/src/vfio_device.rs:312 -- Could not delete VFIO group:
KvmSetDeviceAttr(Error(9))
After some investigation, it appears the KVM device file descriptor used
for removing a VFIO group was already closed. This is coming from the
Rust sequence of Drop, from the DeviceManager all the way down to
VfioDevice.
Because the DeviceManager owns passthrough_device, which is effectively
a KVM device file descriptor, when the DeviceManager is dropped, the
passthrough_device follows, with the effect of closing the KVM device
file descriptor. Problem is, VfioDevice has not been dropped yet and it
still needs a valid KVM device file descriptor.
That's why the simple way to fix this issue coming from Rust dropping
all resources is to make Linux accountable for it by duplicating the
file descriptor. This way, even when the passthrough_device is dropped,
the KVM file descriptor is closed, but a duplicated instance is still
valid and owned by the VfioContainer.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
We turn on that emulation for Windows. Windows does not have KVM's PV
clock, so calling notify_guest_clock_paused results in an error.
Signed-off-by: Wei Liu <liuwe@microsoft.com>
If the user specified a maximum physical bits value through the
`max_phys_bits` option from `--cpus` parameter, the guest CPUID
will be patched accordingly to ensure the guest will find the
right amount of physical bits.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
If the user provided a maximum physical bits value for the vCPUs, the
memory manager will adapt the guest physical address space accordingly
so that devices are not placed further than the specified value.
It's important to note that if the number exceed what is available on
the host, the smaller number will be picked.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
In order to let the user choose maximum address space size, this patch
introduces a new option `max_phys_bits` to the `--cpus` parameter.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The 'GuestAddress::unchecked_add' function has undefined behavior when
an overflow occurs. Its alternative 'checked_add' requires use to handle
the overflow explicitly.
Signed-off-by: Bo Chen <chen.bo@intel.com>
We are now reserving a 256M gap in the guest address space each time
when hotplugging memory with ACPI, which prevents users from hotplugging
memory to the maximum size they requested. We confirm that there is no
need to reserve this gap.
This patch removes the 'reserved gaps'. It also refactors the
'MemoryManager::start_addr' so that it is rounding-up to 128M alignment
when hotplugged memory is allowed with ACPI.
Signed-off-by: Bo Chen <chen.bo@intel.com>
We now try to create a ram region of size 0 when the requested memory
size is the same as current memory size. It results in an error of
`GuestMemoryRegion(Mmap(Os { code: 22, kind: InvalidInput, message:
"Invalid argument" }))`. This error is not meaningful to users and we
should not report it.
Signed-off-by: Bo Chen <chen.bo@intel.com>
This is a new clippy check introduced in 1.47 which requires the use of
the matches!() macro for simple match blocks that return a boolean.
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The OneRegister literally means "one (arbitrary) register". Just call it
"Register" instead. There is no need to inherit KVM's naming scheme in
the hypervisor agnostic code.
Signed-off-by: Wei Liu <liuwe@microsoft.com>
Small patch creating a dedicated `block_io_uring_is_supported()`
function for the non-io_uring case, so that we can simplify the
code in the DeviceManager.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Without the unlink(2) syscall being allowed, Cloud-Hypervisor crashes
when we remove a virtio-vsock device that has been previously added.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Because of the PCI refactoring that happened in the previous commit
d793cc4da3, the ability to fully remove a
PCI device was altered.
The refactoring was correct, but the usage of a generic function to pass
the same reference for both BusDevice, PciDevice and Any + Send + Sync
causes the Arc::ptr_eq() function to behave differently than expected,
as it does not match the references later in the code. That means we
were not able to remove the device reference from the MMIO and/or PIO
buses, which was leading to some bus range overlapping error once we
were trying to add a device again to the previous range that should have
been removed.
Fixes#1802
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Run loop in hypervisor needs a callback mechanism to access resources
like guest memory, mmio, pio etc.
VmmOps trait is introduced here, which is implemented by vmm module.
While handling vcpuexits in run loop, this trait allows hypervisor
module access to the above mentioned resources via callbacks.
Signed-off-by: Praveen Paladugu <prapal@microsoft.com>
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
A new version of vm-memory was released upstream which resulted in some
components pulling in that new version. Update the version number used
to point to the latest version but continue to use our patched version
due to the fix for #1258
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The PMEM support has an option called "discard_writes" which when true
will prevent changes to the device from hitting the backing file. This
is trying to be the equivalent of "readonly" support of the block
device.
Previously the memory of the device was marked as KVM_READONLY. This
resulted in a trap when the guest attempted to write to it resulting a
VM exit (and recently a warning). This has a very detrimental effect on
the performance so instead make "discard_writes" truly CoW by mapping
the memory as `PROT_READ | PROT_WRITE` and using `MAP_PRIVATE` to
establish the CoW mapping.
Fixes: #1795
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
The virtio-balloon change the memory size is asynchronous.
VirtioBalloonConfig.actual of balloon device show current balloon size.
This commit add memory_actual_size to vm.info to show memory actual size.
Signed-off-by: Hui Zhu <teawater@antfin.com>
Write to the exit_evt EventFD which will trigger all the devices and
vCPUs to exit. This is slightly cleaner than just exiting the process as
any temporary files will be removed.
Fixes: #1242
Signed-off-by: Rob Bradford <robert.bradford@intel.com>
This patch adds the missing the `iommu` and `id` option for
`VmAddDevice` in the openApi yaml to respect the internal data structure
in the code base. Also, setting the `id` explicitly for VFIO device
hotplug is required for VFIO device unplug through openAPI calls.
Signed-off-by: Bo Chen <chen.bo@intel.com>
According to openAPI specification [1], the format for `integer` types
can be only `int32` or `int64`, unsigned and 8-bits integers are not
supported.
This patch replaces `uint64` with `int64`, `uint32` with `int32` and
`uint8` with `int32`.
[1]: https://swagger.io/specification/#data-types
Signed-off-by: Julio Montes <julio.montes@intel.com>
MsiInterruptGroup doesn't need to know the internal field names of
InterruptRoute. Introduce two helper functions to eliminate references
to irq_fd. This is done similarly to the enable and disable helper
functions.
Also drop the pub keyword from InterruptRoute fields. It is not needed
anymore.
No functional change.
Signed-off-by: Wei Liu <liuwe@microsoft.com>
According to openAPI specification[1], the format for `integer` types
can be only `int32` or `int64`, unsigned integers are not supported.
This patch replaces `uint64` with `int64`.
[1]: https://swagger.io/specification/#data-types
Signed-off-by: Julio Montes <julio.montes@intel.com>
There is no point in manually dropping the lock for gsi_msi_routes then
instantly grabbing it again in set_gsi_routes.
Make set_gsi_routes take a reference to the routing hashmap instead.
No functional change intended.
Signed-off-by: Wei Liu <liuwe@microsoft.com>
The MTRR feature was missing from the CPUID, which is causing the guest
to ignore the MTRR settings exposed through dedicated MSRs.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Since Cloud-Hypervisor currently support one single PCI bus, we must
reflect this through the MCFG table, as it advertises the first bus and
the last bus available. In this case both are bus 0.
This patch saves quite some time during guest kernel boot, as it
prevents from checking each bus for available devices.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
The states of GIC should be part of the VM states. This commit
enables the AArch64 VM states save/restore by adding save/restore
of GIC states.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
The definition of libc::SYS_ftruncate on AArch64 is different
from that on x86_64. This commit unifies the previously hard-coded
syscall number for AArch64.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
`KVM_GET_REG_LIST` ioctl is needed in save/restore AArch64 vCPU.
Therefore we whitelist this ioctl in seccomp.
Also this commit unifies the `SYS_FTRUNCATE` syscall for x86_64
and AArch64.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
Similarly as the VM booting process, on AArch64 systems,
the vCPUs should be created before the creation of GIC. This
commit refactors the vCPU save/restore code to achieve the
above-mentioned restoring order.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
Since calling `KVM_GET_ONE_REG` before `KVM_VCPU_INIT` will
result in an error: Exec format error (os error 8). This commit
decouples the vCPU init process from `configure_vcpus`. Therefore
in the process of restoring the vCPUs, these vCPUs can be
initialized separately before started.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
The construction of `GICR_TYPER` register will need vCPU states.
Therefore this commit adds methods to extract saved vCPU states
from the cpu manager.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
Unlike x86_64, the "interrupt_controller" in the device manager
for AArch64 is only a `Gic` object that implements the
`InterruptController` to provide the interrupt delivery service.
This is not the real GIC device so that we do not need to save
its states. Also, we do not need to insert it to the device_tree.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
The value of GIC register `GICR_TYPER` is needed in restoring
the GIC states. This commit adds a field in the GIC device struct
and a method to construct its value.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
In AArch64 systems, the state of GIC device can only be
retrieved from `KVM_GET_DEVICE_ATTR` ioctl. Therefore to implement
saving/restoring the GIC states, we need to make sure that the
GIC object (either the file descriptor or the device itself) can
be extracted after the VM is started.
This commit refactors the code of GIC creation by adding a new
field `gic_device_entity` in device manager and methods to set/get
this field. The GIC object can be therefore saved in the device
manager after calling `arch::configure_system`.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
This commit adds a function which allows to save RDIST pending
tables to the guest RAM, as well as unit test case for it.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
This commit adds the unit test cases for getting/setting the GIC
distributor, redistributor and ICC registers.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
Adds 3 more unit test cases for AArch64:
*save_restore_core_regs
*save_restore_system_regs
*get_set_mpstate
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
This commit ports code from firecracker and refactors the existing
AArch64 code as the preparation for implementing save/restore
AArch64 vCPU, including:
1. Modification of `arm64_core_reg` macro to retrive the index of
arm64 core register and implemention of a helper to determine if
a register is a system register.
2. Move some macros and helpers in `arch` crate to the `hypervisor`
crate.
3. Added related unit tests for above functions and macros.
Signed-off-by: Henry Wang <Henry.Wang@arm.com>
Misspellings were identified by https://github.com/marketplace/actions/check-spelling
* Initial corrections suggested by Google Sheets
* Additional corrections by Google Chrome auto-suggest
* Some manual corrections
Signed-off-by: Josh Soref <jsoref@users.noreply.github.com>
virtio-mem device would use 'VIRTIO_MEM_F_ACPI_PXM' to add memory to NUMA
node, which MUST be existed, otherwise it will be assigned to node id 0,
even if user specify different node id.
According ACPI spec about Memory Affinity Structure, system hardware
supports hot-add memory region using 'Hot Pluggable | Enabled' flags.
Signed-off-by: Jiangbo Wu <jiangbo.wu@intel.com>
Use zone.host_numa_node to create memory zone, so that memory zone
can apply memory policy in according with host numa node ID
Signed-off-by: Jiangbo Wu <jiangbo.wu@intel.com>
If after the creation of the self-spawned backend, the VMM cannot create
the corresponding vhost-user frontend, the VMM must kill the freshly
spawned process in order to ensure the error propagation can happen.
In case the child process would still be around, the VMM cannot return
the error as it waits onto the child to terminate.
This should help us identify when self-spawned failures are caused by a
connection being refused between the VMM and the backend.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
When the VMM is terminated by receiving a SIGTERM signal, the signal
handler thread must be able to invoke ioctl(TCGETS) and ioctl(TCSETS)
without error.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Based on all the preparatory work achieved through previous commits,
this patch updates the 'hotplugged_size' field for both MemoryConfig and
MemoryZoneConfig structures when either the whole memory is resized, or
simply when a memory zone is resized.
This fixes the lack of support for rebooting a VM with the right amount
of memory plugged in.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Adding a new field to VirtioMemZone structure, as it lets us associate
with a particular virtio-mem region the amount of memory that should be
plugged in at boot.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
This patch simplifies the code as we have one single Option for the
VirtioMemZone. This also prepares for storing additional information
related to the virtio-mem region.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>
Add the new option 'hotplugged_size' to both --memory-zone and --memory
parameters so that we can let the user specify a certain amount of
memory being plugged at boot.
This is also part of making sure we can store the virtio-mem size over a
reboot of the VM.
Signed-off-by: Sebastien Boeuf <sebastien.boeuf@intel.com>